claesenm / approxsvm Goto Github PK

*** ApproxSVM ***
Approximating nonlinear SVM models with RBF kernel.

This project is a C++ implementation of a second-order Maclaurin series 
approximation of LIBSVM models using an RBF kernel. This approximation
greatly increases prediction speed for models with many support vectors
but in few dimensions. 

For mathematical details, please refer to the associated publication at:
  ftp://ftp.esat.kuleuven.be/pub/SISTA/claesenm/reports/14-26.pdf
  
In this project we only provide the implementation of the approximation. 
For reference, a script and data to execute most of the experiments 
reported in the manuscript listed above using the tools in this project 
are available at the following location (about 30 MB in size):
  http://homes.esat.kuleuven.be/~claesenm/approxsvm/examples.tar.gz
  
If you use this software, please acknowledge it in publications.

*******************************************************************

COMPILATION

Depending on what facilities are available on your platform, several
options exist for computations. The configure script detects presence
of BLAS routines on your platform and uses them if possible. If not
available, a naive (slow) fallback implementation is used instead. If
you have ATLAS on your system, make sure to enable it in the configure
stage using --with-atlas and include the path to the ATLAS library
in LDFLAGS, for example using LDFLAGS="-L/usr/local/atlas/lib".

It is strongly advised to allow the use of vector instructions using
the following compiler flag: "-march=native".

For a typical build, issue the following commands:
  ./configure CXXFLAGS="-O3 -march=native" --disable-shared
  make
  sudo make install

If you get errors about loading shared libraries, issue this command:
  sudo ldconfig

*******************************************************************

USING THE TOOLS

This software package contains the following tools:
  approx-svm: used to approximate a given LIBSVM model with RBF kernel
  approx-predict: used to make predictions with an approximated model
  approx-analyse: analyse data set and determine maximum value for 
    gamma for which a model can be reliably approximated
  approx-xval: performs k-fold cross-validation using approximated
    models instead of exact LIBSVM models (train-approx-predict)
  
For an overview of command line arguments, please execute either tool
without arguments or with "--help".

*******************************************************************

CONTACT

Please report bugs or suggestions to [email protected]. 
The latest version of this software is freely available at:
  https://github.com/claesenm/approxsvm